01Omnigent. Databricks open-sources the missing governance layer for enterprise AI agents.
02This week, in brief. Spatial SQL GA, Zerobus ingest, Lakebase branching, and named enterprise AI wins from Data + AI Summit 2026.
03From Brickster.ai: Your curated Databricks digest for the week.
01
🤖 Omnigent
Your agents are running wild. Databricks just built the fence.
Omnigent is open source, Apache 2.0, and solves the exact problem your engineering org is quietly living with.
On June 13, 2026, two days before the Data + AI Summit kicked off, Databricks quietly dropped something significant: Omnigent, an open-source "meta-harness" for AI agents. Not a new model. Not a new platform. Something more pragmatic: a governance and collaboration layer that sits above the agent tools your teams already use.
It works with Claude Code, Codex, Pi, and any custom agent you have built. The pitch is simple: stop managing each agent in its own silo, and stop relying on prompts to enforce your security and cost policies.
Databricks built this because they ran into the problem themselves, across their 5,000+ member engineering team. When they rolled out coding agents at scale, engineers were jumping between Claude Code, Codex, and Pi with no shared sessions, no unified policies, and no way to collaborate. Omnigent is the solution they built internally, now open-sourced.
🔥 The problem it solves
The agent fragmentation trap
Most enterprise engineering teams today run multiple AI agents, with different tools for different purposes, each with its own CLI, its own session, its own security model. The result is fragmentation: context gets copy-pasted between tools, policies differ per team, sensitive credentials get exposed to agents that should not see them, and when something goes wrong there is no unified audit trail.
That fragmentation is a big part of why enterprise AI agents stay stuck in pilot mode. The governance story does not hold up at production scale, so the rollout stalls before it ever reaches the whole org.
The core insight behind Omnigent: whichever agent harness you use, the interface is always the same. Messages and files go in, text streams and tool calls come out. Omnigent standardises that interface and adds a policy layer on top.
⚙️ What Omnigent gives you
Pillar 1
🔀 Composition: stop rewriting, start composing
Swap between Claude Code, Codex, Pi, or a custom YAML-defined agent with a single-line change. Run multiple agents in parallel on the same shared session. Two built-in agents ship out of the box: Polly (a coding orchestrator) and Debby (a model-debate agent that pits two models against each other to compare answers). Build your own in YAML.
Pillar 2
🔒 Control: policies, not prompts
This is the part enterprises will care about. Omnigent enforces governance at the orchestration layer, not via fragile prompt engineering. Specific controls include:
Spend caps: auto-pause an agent the moment it hits a cost threshold, and require human approval to resume
Risk-based escalation: require a human approval gate on specific risky tool calls (for example, git push after modifying dependencies)
OS-level sandbox (Omnibox): restrict filesystem and network access, hide credentials from the agent, and broker access on its behalf, so the agent never directly sees your API keys
Model routing: automatically route different task types to different models
Pillar 3
🤝 Collaboration: share live sessions, not screenshots
Share a live agent session via URL. Teammates join with full history. They can leave inline comments on specific agent actions (like a code review, but for what the agent did), and co-steer the agent in real time. Access from terminal, web UI, native macOS app, or mobile, all interfaces staying in sync across messages, terminals, sub-agents, and file workspaces.
🏗️ How it works
Omnigent has two components. The Runner wraps any agent in a sandboxed, uniform session, treating each harness as an interchangeable worker. The Server manages policies, stores session history, and exposes every session over the terminal, a web UI, a native app, mobile, and a REST API.
Omnigent Server
Adds policies, shared history, and session exposure
Terminal
Web UI
Native App
Mobile
REST API
Omnigent Runner
Sandboxed, uniform session wrapping any agent
Claude Code
Codex
Pi
Custom YAML
Install is a single command (also on pip, uv, and Homebrew, with a macOS native app alongside the CLI):
curl -fsSL https://omnigent.ai/install.sh | sh
Alpha
Status
Jun 13
Released
Apache 2
License
Open
Source
🔭 What this means for your data & AI org
If your teams are already using coding agents, even experimentally, you are likely already in the fragmentation trap. Different squads using different tools, no shared visibility, no consistent cost controls, security teams unsure what data the agents are actually touching.
Omnigent is worth watching for three reasons if you lead a data or AI org:
1. Vendor neutrality. Apache 2.0 means no lock-in. You keep Claude Code and Codex. Omnigent just orchestrates them.
2. A credible governance story. The sandbox and credential brokering speak to the security-perimeter questions your CISO is already asking about agentic AI. It is alpha, so treat it as something to pilot, not a compliance guarantee.
3. It is Alpha, so get in early. The feedback loop is open on Discord. If you deploy agents at scale, this is a chance to influence direction before it solidifies.
Beyond Omnigent, Data + AI Summit 2026 ran three clear threads through the brickster.ai archive this week. Full breakdown at brickster.ai/digest.
Open formats
The lakehouse keeps opening up
Spatial SQL hit GA with AI/BI Maps, Delta Sharing, and Iceberg v3, and a new storage-ecosystem push lets Unity Catalog govern data that never leaves your own data center. The Summit throughline: open formats and governance reaching wherever the data already lives.
Zerobus Ingest showed petabyte-scale streaming with a literal Milky Way dataset, while the Lakebase database-branching series wrapped with the production playbook. The plumbing under the AI demos is getting a lot sturdier.
The customer stories did the talking this week. Mercedes-Benz Korea built a trusted "Talk to Data" layer on semantic models, and Ecolab rebuilt its retail intelligence on Databricks and Anthropic Claude. Both are the "we shipped it" proof points the keynotes lean on.
Databricks shipped dozens of announcements this week. The home page now shows them as they land, ranked for your role.
brickster.ai pulls news, releases, videos, GitHub projects, and community Q&A from across the ecosystem and refreshes them every morning. During a week like this one, that is the difference between reading the signal and drowning in the keynote recap. Pick your role and the whole board reorders around what you actually work on.
Or skip the reading entirely: the assistant answers in plain English with citations straight from the archive, so you can ask what shipped at Summit and get a sourced reply.
